Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization

Xudong Wang; Chaoning Zhang; Jiaquan Zhang; Chenghao Li; Qigan Sun; Sung-Ho Bae; Peng Wang; Ning Xie

cs.AIMar 13, 2026

Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization

Authors: Xudong Wang, Chaoning Zhang, Jiaquan Zhang, Chenghao Li, Qigan Sun, Sung-Ho Bae, Peng Wang, Ning Xie,

Organizations: School of Computing, Kyung Hee University, Yongin-si 17104, South Korea · School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China · School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 610054, China · School of Computer Science and Technology, Tongji University, Shanghai 200092, China

Abstract

Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.

Figures & tables

Fig. 1: Overview of the AMRO-S routing mechanism. Tasks are routed through three stages, collection, analysis, and solution, via probabilistic path sampling guided by dynamic pheromone signals. After execution, high-quality paths receive reinforced pheromones, increasing their selection likelihood.

Fig. 2: Architecture of AMRO-S. (a) Offline construction of layered graph $G=(V,E)$ and pheromone specialists. (b) Online routing via SFT-SLM weights $w(q)$ across three stages, where nodes represent (LLM, Method, Role) instances. (c) Asynchronous evolution using LLM-Judge quality gating ( $g\in\{0,1\}$ ) for background pheromone reinforcement without serving overhead.

Reference Models
Method	LLM	Mul.	Rout.	MMLU	GSM8K	MATH	HumanEval	MBPP	Avg.
Vanilla (Ref)	GPT-4o	N	N	88.7	96.1	76.6	90.2	87.2	87.76
Vanilla (Ref)	Claude-3.5-Sonnet	N	N	88.3	96.4	78.0	93.7	89.1	89.10
Single-Agent Baselines
Vanilla	GPT-4o-mini	N	N	77.81	93.17	66.09	85.71	72.2	79.00
Vanilla	Gemini-1.5-flash	N	N	80.04	92.67	74.39	82.61	73.0	80.54

TABLE I: Performance comparison on five benchmarks. Mul. and Rout. denote Multi-Agent and Dynamic Routing, respectively. The LLM Pool* comprises cost-effective models (GPT-4o-mini, Gemini-1.5-flash, Claude-3.5-haiku, Llama-3.1-70b).

Framework	Dataset	Model	Original		+MasRouter		+AMRO-S (Ours)
Framework	Dataset	Model	Acc.	Cost	Acc.	Cost	Acc.	Cost
MacNet	MMLU	GPT	82.98%	7.81	83.25%	7.65	83.50%	7.50
	MMLU	Gemini	81.74%	8.90	81.95%	8.65	82.10%	8.40
	HumanEval	GPT	86.82%	0.49	87.20%	0.48	87.50%	0.47
	HumanEval	Gemini	88.72%	0.53	88.85%	0.52	89.00%	0.50
	GSM8K	GPT	94.69%	2.14	94.85%	2.08	95.00%	2.00

TABLE II: Adaptability and Cost-Efficiency Evaluation on Existing Multi-Agent Frameworks. We integrate the classifier-based MasRouter and our pheromone-based AMRO-S into three representative frameworks (MacNet, GPTSwarm, HEnRY). GPT: GPT-4o-mini; Gemini: Gemini-1.5-flash.

ID	Setting	Router Backbone	Strategy	Mul.	Rout.	GSM8K	MATH	MMLU	HumanEval	MBPP	Avg.
Constituents (Single-Agent Baselines)
(A)	Single	Gpt-4o-mini	N/A	$\times$	$\times$	93.17	66.09	77.81	85.71	72.20	79.00
(B)	Single	Gemini-1.5-flash	N/A	$\times$	$\times$	92.67	74.39	80.04	82.61	73.00	80.54
(C)	Single	Claude-3.5-haiku	N/A	$\times$	$\times$	91.80	68.20	78.50	86.40	74.50	79.88
(D)	Single	Llama-3.1-70b	N/A	$\times$	$\times$	94.10	68.00	82.30	80.50	71.80	79.34
Router and Strategy Ablation

TABLE III: Ablation Study on Router Components. Analysis of single-agent baselines versus different router backbones and strategies.

Model	SFT	Math	Code	General	Avg.
Llama-3.2-1B-Instruct	$\times$	78.50%	82.10%	85.40%	82.00%
Qwen2.5-1.5B	$\times$	84.20%	88.50%	89.10%	87.26%
GPT-4o-mini	$\times$	95.20%	96.10%	96.20%	95.83%
Qwen2.5-1.5B	$\checkmark$	97.90%	98.20%	97.50%	97.86%
Llama-3.2-1B-Instruct	$\checkmark$	98.10%	97.90%	97.80%	97.93%

TABLE IV: SLM Intent Recognition Accuracy. Comparison of zero-shot baselines and fine-tuned models (SFT) across different intents.

Processes	20	50	100	200	500	1000
AMRO-S Efficiency Metrics
Time (s)	3849.60	2430.40	1863.75	1382.90	1062.30	823.21
Time (min)	64.16	40.51	31.06	23.05	17.71	13.72
Speedup	1.0 $\times$	1.6 $\times$	2.1 $\times$	2.8 $\times$	3.6 $\times$	4.7 $\times$
Accuracy Comparison
WRR(Baseline)	96.00%	95.80%	95.20%	93.50%	91.50%	88.20%

TABLE V: Stress test under varying concurrency. “Processes” denotes the number of parallel workers; speedup is relative to 20 processes.

Fig. 3: Converged pheromone specialists of AMRO-S for three domains: mathematical reasoning $T_{math}$ , code generation $T_{code}$ , and general reasoning $T_{gen}$ . Color intensity indicates the learned routing preference, where deeper teal denotes stronger preference and lighter tones denote weaker preference.

LLM Agent Routing

LLM agent routing selects models, tools, skills, or collaboration workflows for each task or step to balance capability, cost, latency, and reliability.

See all 99 papers

Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing
Jun 29, 2026Dvir Alsheich, Adar Peleg, Ben Hagag +3Adversarial Attacks on LLMs Agent Evaluation
1Technion - Israel Institute of Technology · 2Carnegie Mellon University
Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models
May 15, 2026Xingjian Wu, Junkai Lu, Siyu Yan +4Multi-Agent Coordination Multi-Agent Systems
East China Normal University
Optimal-Agent-Selection: State-Aware Routing Framework for Efficient Multi-Agent Collaboration
Nov 4, 2025Jingbo Wang, Sendong Zhao, Haochun Wang +2Multi-Agent System Optimization Multi-Agent Collaboration
Research Center for Social Computing and Information Retrieval,Harbin Institute of Technology, China

Jun 29, 2026J/K move · Enter open · S save

Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing

Dvir Alsheich, Adar Peleg, Ben Hagag, Rom Himelstein+2

1Technion - Israel Institute of Technology · 2Carnegie Mellon University

Adversarial Attacks on LLMs Agent Evaluation Multi-Agent System Security LLM Agent Routing

Open paper AskPDF arXiv

The rapid integration of Large Language Models (LLMs) has driven the evolution of Multi-Agent Systems (MAS), where specialized agents collaborate to execute complex workflows. Effective orchestration in these environments requires robust routing mechanisms to efficiently allocate tasks to the most suitable agent. However, existing routers fundamentally rely on unverified proxies, ranging from textual self-descriptions to static surrogate representations, to gauge an agent's competence. This reliance on non-empirical data creates a critical gap between an agent's projected profile and its actual operational capabilities, introducing severe security vulnerabilities. Malicious agents can easily misrepresent their proficiencies or harbor covert backdoors that evade both standard external analysis and static representation-learning techniques. In this work, we introduce ANTAP (Automatic Non-Textual Agent Picker), an evaluation-driven routing architecture that discards indirect proxies in favor of active capability testing. By dynamically querying agents to ascertain their true competencies empirically, ANTAP distills performance into fixed behavioral operators within a shared semantic space. At inference time, routing is performed via a purely non-textual algebraic projection, establishing a "linguistic firewall" that renders metadata-based attacks inexpressible. In our experiments, ANTAP achieves near-zero ASR against description-based injection attacks, compared to 67.3% and above for the description-based router baseline. Against adaptive embedding attacks, ANTAP achieves substantially lower ASR than the embedding-based baseline, with a 20% reduction, while remaining resilient to description manipulation by design.

Abstract

LLM Agent Routing

Explore similar work

Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing

Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models

Optimal-Agent-Selection: State-Aware Routing Framework for Efficient Multi-Agent Collaboration